Comparison of hot rolled steel mechanical property prediction models using linear multiple regression, non-linear multiple regression and non-linear artificial neural networks

Abstract

In the manufacture of rolled steel from a hot strip mill, the final mechanical properties, such as yield strength, ultimate tensile strength and elongation to fracture, are important requirements specified by the customer. The use of mathematical modelling techniques such as multiple regression analysis, or computational developments such as artificial neural networks, can result in the creation of acceptably accurate predictive models. However, the accuracy of any predictive model will depend on the quality of data used in its creation, and thus a brief statistical analysis of the mechanical property data used for model development is discussed. In the present paper a comparison of the application of linear multiple regression, non-linear multiple regression and non-linear neural networks is made for various steel families using data taken from the Corus Port Talbot hot strip mill. A statistical summary of their relative predictive errors is given, and although all three are comparable, the non-linear, black box approach of a suitably structured neural network provides overall more accurate predictive models than the use of linear or non-linear multiple regression.

Keywords:

• INDIVIDUAL AND MOVING RANGE
• LINEAR MULTIPLE REGRESSION
• NON-LINEAR MULTIPLE REGRESSION
• NEURAL NETWORKS
• YIELD STRENGTH
• ULTIMATE TENSILE STRENGTH
• ELONGATION TO FRACTURE